Predicting rental listing interest

Data: rental listings of aparments in NY City

Goal: predict level of interests

Some observations/questions raised during data exploration (would need to check in real-world use-case):

• Data is quite homogenous as restricted as apartments in NY city. This is like building a model for each local area, which may work well if there is a lot of data.

• Is target normalised to the same listing period of length (e.g. number of days or weeks)? I suspect not as there is some correlation between listing id and interest level.

• Use at least 2-year of historical data to build model to account for seasonal effects (e.g. school holidays, vacation)

• Should also use out of time sampling and testing

Target

Target distribution shows imbalanced classes.

table(d$interest_level)

## 
##   high    low medium 
##   3839  34284  11229

Interest level is quantitative and we can encode the targets as numerical values and build a regression model. Future work.

Exploratory analysis

sapply(d, class)

##       bathrooms        bedrooms     building_id         created 
##       "numeric"       "integer"     "character"     "character" 
##     description display_address        features        latitude 
##     "character"     "character"          "list"       "numeric" 
##      listing_id       longitude      manager_id          photos 
##       "integer"       "numeric"     "character"          "list" 
##           price  street_address  interest_level 
##       "integer"     "character"     "character"

• Train and test period?
• Identifiers?
• Missing values? Outliers?
• Additional features?
• Features vs. features correlations?
• Features vs. target correlations?

Train and test periods

Train and test in the same periods of 3 months, so there could be information leak due to seasonal effects. Proportion of listings in each month is same in both train and test.

min(d$created)

## [1] "2016-04-01 22:12:41"

max(d$created)

## [1] "2016-06-29 21:41:47"

min(d_test$created)

## [1] "2016-04-01 22:23:31"

max(d_test$created)

## [1] "2016-06-29 21:55:35"

d <- mutate(d, month_created = month(as_datetime(created)))
d_test <- mutate(d_test, month_created = month(as_datetime(created)))
table(d$month_created)

## 
##     4     5     6 
## 16411 16626 16315

table(d_test$month_created)

## 
##     4     5     6 
## 24879 24987 24793

Identifier features

Listing IDs are identifiers so remove.

d <- select(d, -listing_id)

Missing values and outliers?

There are missing values for latitude and longitude (entered as 0), and outliers for prices. Leave them in for now as tree-based model can handle such cases

summary(select(d, bathrooms, bedrooms, latitude, longitude, price))

##    bathrooms         bedrooms        latitude       longitude      
##  Min.   : 0.000   Min.   :0.000   Min.   : 0.00   Min.   :-118.27  
##  1st Qu.: 1.000   1st Qu.:1.000   1st Qu.:40.73   1st Qu.: -73.99  
##  Median : 1.000   Median :1.000   Median :40.75   Median : -73.98  
##  Mean   : 1.212   Mean   :1.542   Mean   :40.74   Mean   : -73.96  
##  3rd Qu.: 1.000   3rd Qu.:2.000   3rd Qu.:40.77   3rd Qu.: -73.95  
##  Max.   :10.000   Max.   :8.000   Max.   :44.88   Max.   :   0.00  
##      price        
##  Min.   :     43  
##  1st Qu.:   2500  
##  Median :   3150  
##  Mean   :   3830  
##  3rd Qu.:   4100  
##  Max.   :4490000

sum(d$latitude == 0)

## [1] 12

sum(d$longitude == 0)

## [1] 12

Categorical features

Intuitively we expect manager, buildings, and street address all to be predictive of the target. But they are categorical features with many values. Using them directly may lead to complex model that can overfit, so we try to convert them into vector of real values. However we must be careful to not leak information from train to validation during cross-validation, especially for managers / building with few listings. Use threshold to determinte when to use.

length(unique(d$building_id))

## [1] 7585

length(unique(d$manager_id))

## [1] 3481

length(unique(d$display_address))

## [1] 8630

length(unique(d$street_address))

## [1] 15358

In fact, we see that they all exhibit heavy tail distribution, meaning a top few instances have many listings and the majority have few listings.

inspect_categorical_var(d, 'building_id')

inspect_categorical_var(d, 'manager_id')

inspect_categorical_var(d, 'display_address')

Manager

Build features for each manager in terms of the interest levels to their properties. We see that managers vary in their listing performance, for example

qplot(x = mngr_prop_high, data=manager_features, geom='histogram', binwidth=0.1)

qplot(x = mngr_prop_low, data=manager_features, geom='histogram', binwidth=0.1)

Display address

We expect property location to contribute to interest level. This information is contained in lat and long, but we can also leverage display address to gauge the interest at street level. But to aggregate on street level, we need to see if there are enough houses on street.

street_features <- dplyr::count(d, display_address)
summary(street_features$n)

##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##   1.000   1.000   1.000   5.719   3.000 438.000

print('top 10% percentile street')

## [1] "top 10% percentile street"

quantile(street_features$n, 0.9)

## 90% 
##  10

So 10% of streets have at least 10 houses, we may build features for these top streets.

Geo-spatial

We zoom in in the centre.

d_zoom <- d %>% filter(longitude > -74.5 & longitude < -73.5 & latitude > 40.5 & latitude < 41)
ggplotly(ggplot(d_zoom, aes(longitude, latitude)) + geom_jitter(aes(color = interest_level), alpha=0.5, size=0.2))

Looking at density on each dimension we see that there are relatively higher concentration of high interest properties near the edge.

ggplot(d_zoom, aes(x = latitude, color=interest_level)) + geom_density()

ggplot(d_zoom, aes(x = longitude, color=interest_level)) + geom_density()

Can also look at medium and high only, but same information as prev plots.

ggplotly(filter(d_zoom, interest_level != 'low') %>% ggplot(aes(longitude, latitude)) + geom_jitter(aes(color = interest_level), alpha=0.5, size=0.5))

Augmenting features

We augement features with manager, street, and building features.

Correlations

Interest levels are ordinal, so convert to -2, 0, 2 and plot correlations. The newly features correlate much stronger to the target than original features.

d2 <- mutate(d_train, interest = ifelse(interest_level == 'high', 2, ifelse(interest_level == 'medium', 1, 0)))
numeric_features <- names(d2)[which(sapply(d2, is.numeric) | sapply(d2, is.integer))]
cor(select(d2, one_of(numeric_features[1:10]), interest),
    method = "spearman", use="pairwise.complete.obs") %>% corrplot(method = "number")

cor(select(d2, one_of(numeric_features[11:20]), interest),
    method = "spearman", use="pairwise.complete.obs") %>% corrplot(method = "number")

cor(select(d2, one_of(numeric_features[21:39]), interest),
    method = "spearman", , use="pairwise.complete.obs") %>% corrplot(method = "number")

## Warning in one_of(numeric_features[21:39]): Unknown variables: `NA`

Building model - first pass

Let’s build a first pass model using numerical features only. We can deal with text features later and ensemble the models.

Since data is very imbalanced, we should use xgboost with case weights

The CV error:

params <- list(eta = 0.2, gamma = 0, max_depth = 6, min_child_weight = 2,
               subsample = 1, colsample_bytree = 0.7,
               num_class = 3, objective = "multi:softprob", eval_metric = "mlogloss")
xgb_d_train <- xgb.DMatrix(data.matrix(select(d_train, -interest_level)),
                           label=to_numeric(d_train$interest_level),
                           #weight=case_weights(d_train$interest_level),
                           missing=NA)
xgb_d_val <- xgb.DMatrix(as.matrix(select(d_val, -interest_level)),
                         label=to_numeric(d_val$interest_level), missing=NA)

xgb_cv <- xgb.cv(params, xgb_d_train, nfold = 5, nrounds = 200, early.stop.round = 3, nthread = 4)

## [0]  train-mlogloss:0.975658+0.000430    test-mlogloss:0.978276+0.002567
## [1]  train-mlogloss:0.891308+0.001711    test-mlogloss:0.896319+0.003286
## [2]  train-mlogloss:0.826890+0.002722    test-mlogloss:0.834186+0.004270
## [3]  train-mlogloss:0.778011+0.002824    test-mlogloss:0.787518+0.005751
## [4]  train-mlogloss:0.740331+0.002795    test-mlogloss:0.751918+0.007003
## [5]  train-mlogloss:0.709880+0.003059    test-mlogloss:0.723541+0.008167
## [6]  train-mlogloss:0.684465+0.002456    test-mlogloss:0.700254+0.008787
## [7]  train-mlogloss:0.665196+0.001879    test-mlogloss:0.682704+0.009882
## [8]  train-mlogloss:0.648491+0.002415    test-mlogloss:0.667708+0.010396
## [9]  train-mlogloss:0.635054+0.002453    test-mlogloss:0.655909+0.010157
## [10] train-mlogloss:0.623336+0.001712    test-mlogloss:0.645819+0.011107
## [11] train-mlogloss:0.614134+0.001842    test-mlogloss:0.638026+0.011630
## [12] train-mlogloss:0.605720+0.002608    test-mlogloss:0.631306+0.011715
## [13] train-mlogloss:0.598375+0.002192    test-mlogloss:0.625359+0.011649
## [14] train-mlogloss:0.591140+0.002253    test-mlogloss:0.619835+0.012330
## [15] train-mlogloss:0.585345+0.001637    test-mlogloss:0.615354+0.012511
## [16] train-mlogloss:0.579080+0.002636    test-mlogloss:0.610713+0.011939
## [17] train-mlogloss:0.573491+0.002345    test-mlogloss:0.606533+0.012236
## [18] train-mlogloss:0.568807+0.002532    test-mlogloss:0.603161+0.012168
## [19] train-mlogloss:0.564464+0.002699    test-mlogloss:0.600194+0.012240
## [20] train-mlogloss:0.560930+0.002898    test-mlogloss:0.597806+0.012083
## [21] train-mlogloss:0.557446+0.003100    test-mlogloss:0.595803+0.012002
## [22] train-mlogloss:0.554403+0.003009    test-mlogloss:0.593935+0.012145
## [23] train-mlogloss:0.551447+0.003563    test-mlogloss:0.592240+0.011909
## [24] train-mlogloss:0.548464+0.003755    test-mlogloss:0.590671+0.012105
## [25] train-mlogloss:0.545571+0.004182    test-mlogloss:0.589393+0.011777
## [26] train-mlogloss:0.543044+0.004346    test-mlogloss:0.588047+0.011581
## [27] train-mlogloss:0.540234+0.003755    test-mlogloss:0.586630+0.011844
## [28] train-mlogloss:0.537753+0.003585    test-mlogloss:0.585549+0.011968
## [29] train-mlogloss:0.535315+0.003391    test-mlogloss:0.584541+0.012197
## [30] train-mlogloss:0.532663+0.003374    test-mlogloss:0.583288+0.012317
## [31] train-mlogloss:0.530812+0.003740    test-mlogloss:0.582530+0.012134
## [32] train-mlogloss:0.528689+0.003750    test-mlogloss:0.581529+0.012394
## [33] train-mlogloss:0.526245+0.003710    test-mlogloss:0.580697+0.012293
## [34] train-mlogloss:0.524128+0.003750    test-mlogloss:0.579800+0.012347
## [35] train-mlogloss:0.522219+0.003444    test-mlogloss:0.579116+0.012306
## [36] train-mlogloss:0.520660+0.003472    test-mlogloss:0.578578+0.012306
## [37] train-mlogloss:0.519154+0.003623    test-mlogloss:0.578059+0.012303
## [38] train-mlogloss:0.517316+0.003468    test-mlogloss:0.577508+0.012354
## [39] train-mlogloss:0.515643+0.003364    test-mlogloss:0.576867+0.012402
## [40] train-mlogloss:0.514125+0.003555    test-mlogloss:0.576361+0.012395
## [41] train-mlogloss:0.512855+0.003614    test-mlogloss:0.575968+0.012409
## [42] train-mlogloss:0.510892+0.003478    test-mlogloss:0.575567+0.012591
## [43] train-mlogloss:0.509356+0.003455    test-mlogloss:0.575082+0.012532
## [44] train-mlogloss:0.508112+0.003257    test-mlogloss:0.574789+0.012644
## [45] train-mlogloss:0.506126+0.003169    test-mlogloss:0.574399+0.012756
## [46] train-mlogloss:0.504673+0.003423    test-mlogloss:0.573898+0.012839
## [47] train-mlogloss:0.503133+0.003507    test-mlogloss:0.573637+0.012888
## [48] train-mlogloss:0.501522+0.003603    test-mlogloss:0.573297+0.012886
## [49] train-mlogloss:0.499806+0.003924    test-mlogloss:0.572882+0.012953
## [50] train-mlogloss:0.498512+0.003668    test-mlogloss:0.572587+0.013059
## [51] train-mlogloss:0.497459+0.003838    test-mlogloss:0.572391+0.013004
## [52] train-mlogloss:0.495963+0.003772    test-mlogloss:0.571982+0.013247
## [53] train-mlogloss:0.494604+0.004112    test-mlogloss:0.571800+0.013221
## [54] train-mlogloss:0.493227+0.004354    test-mlogloss:0.571670+0.013281
## [55] train-mlogloss:0.491718+0.004465    test-mlogloss:0.571347+0.013369
## [56] train-mlogloss:0.490451+0.004008    test-mlogloss:0.571079+0.013434
## [57] train-mlogloss:0.488950+0.003803    test-mlogloss:0.570800+0.013453
## [58] train-mlogloss:0.487356+0.003738    test-mlogloss:0.570614+0.013497
## [59] train-mlogloss:0.486000+0.003789    test-mlogloss:0.570397+0.013585
## [60] train-mlogloss:0.484655+0.003775    test-mlogloss:0.570149+0.013511
## [61] train-mlogloss:0.483656+0.003801    test-mlogloss:0.569981+0.013575
## [62] train-mlogloss:0.482649+0.003641    test-mlogloss:0.569836+0.013569
## [63] train-mlogloss:0.481344+0.003744    test-mlogloss:0.569783+0.013715
## [64] train-mlogloss:0.479968+0.003893    test-mlogloss:0.569673+0.013719
## [65] train-mlogloss:0.478737+0.003780    test-mlogloss:0.569546+0.013598
## [66] train-mlogloss:0.477278+0.003814    test-mlogloss:0.569390+0.013506
## [67] train-mlogloss:0.475840+0.003847    test-mlogloss:0.569206+0.013539
## [68] train-mlogloss:0.474472+0.003706    test-mlogloss:0.568915+0.013648
## [69] train-mlogloss:0.473138+0.003746    test-mlogloss:0.568809+0.013734
## [70] train-mlogloss:0.472044+0.003487    test-mlogloss:0.568647+0.013801
## [71] train-mlogloss:0.470815+0.003778    test-mlogloss:0.568539+0.013922
## [72] train-mlogloss:0.469652+0.003934    test-mlogloss:0.568433+0.013879
## [73] train-mlogloss:0.468603+0.004175    test-mlogloss:0.568296+0.013916
## [74] train-mlogloss:0.467411+0.004046    test-mlogloss:0.568290+0.014004
## [75] train-mlogloss:0.466149+0.003956    test-mlogloss:0.568114+0.013943
## [76] train-mlogloss:0.465339+0.003658    test-mlogloss:0.568005+0.013991
## [77] train-mlogloss:0.463978+0.003922    test-mlogloss:0.567910+0.014069
## [78] train-mlogloss:0.462658+0.004264    test-mlogloss:0.567847+0.014041
## [79] train-mlogloss:0.461700+0.004212    test-mlogloss:0.567769+0.014106
## [80] train-mlogloss:0.460921+0.004252    test-mlogloss:0.567733+0.014124
## [81] train-mlogloss:0.460093+0.004161    test-mlogloss:0.567685+0.014119
## [82] train-mlogloss:0.459198+0.004137    test-mlogloss:0.567670+0.014064
## [83] train-mlogloss:0.458149+0.004068    test-mlogloss:0.567595+0.014066
## [84] train-mlogloss:0.457133+0.004007    test-mlogloss:0.567546+0.014109
## [85] train-mlogloss:0.456097+0.003927    test-mlogloss:0.567575+0.014126
## [86] train-mlogloss:0.455198+0.003818    test-mlogloss:0.567499+0.014133
## [87] train-mlogloss:0.454256+0.003832    test-mlogloss:0.567522+0.014231
## [88] train-mlogloss:0.453194+0.003850    test-mlogloss:0.567455+0.014160
## [89] train-mlogloss:0.451930+0.003552    test-mlogloss:0.567428+0.014235
## [90] train-mlogloss:0.450709+0.003518    test-mlogloss:0.567446+0.014219
## [91] train-mlogloss:0.449615+0.003564    test-mlogloss:0.567445+0.014346
## [92] train-mlogloss:0.448628+0.003476    test-mlogloss:0.567423+0.014340
## [93] train-mlogloss:0.447457+0.003516    test-mlogloss:0.567432+0.014323
## [94] train-mlogloss:0.446311+0.003508    test-mlogloss:0.567525+0.014353
## [95] train-mlogloss:0.445222+0.003437    test-mlogloss:0.567527+0.014361
## Stopping. Best iteration: 93

xgb_model <- xgb.train(params, xgb_d_train, 
                       nrounds = 500, early.stop.round = 3, nthread = 4,
                       watchlist = list(val = xgb_d_val))

## [0]  val-mlogloss:0.979630
## [1]  val-mlogloss:0.898859
## [2]  val-mlogloss:0.840962
## [3]  val-mlogloss:0.798048
## [4]  val-mlogloss:0.761406
## [5]  val-mlogloss:0.732785
## [6]  val-mlogloss:0.711867
## [7]  val-mlogloss:0.693052
## [8]  val-mlogloss:0.679435
## [9]  val-mlogloss:0.669895
## [10] val-mlogloss:0.658097
## [11] val-mlogloss:0.651483
## [12] val-mlogloss:0.643463
## [13] val-mlogloss:0.635880
## [14] val-mlogloss:0.631997
## [15] val-mlogloss:0.627317
## [16] val-mlogloss:0.624271
## [17] val-mlogloss:0.620170
## [18] val-mlogloss:0.617388
## [19] val-mlogloss:0.615234
## [20] val-mlogloss:0.614006
## [21] val-mlogloss:0.609879
## [22] val-mlogloss:0.608206
## [23] val-mlogloss:0.606996
## [24] val-mlogloss:0.605774
## [25] val-mlogloss:0.604661
## [26] val-mlogloss:0.603672
## [27] val-mlogloss:0.602863
## [28] val-mlogloss:0.601464
## [29] val-mlogloss:0.600711
## [30] val-mlogloss:0.599235
## [31] val-mlogloss:0.598486
## [32] val-mlogloss:0.597323
## [33] val-mlogloss:0.596512
## [34] val-mlogloss:0.595787
## [35] val-mlogloss:0.595490
## [36] val-mlogloss:0.594993
## [37] val-mlogloss:0.594946
## [38] val-mlogloss:0.594590
## [39] val-mlogloss:0.594271
## [40] val-mlogloss:0.593994
## [41] val-mlogloss:0.593954
## [42] val-mlogloss:0.593739
## [43] val-mlogloss:0.593048
## [44] val-mlogloss:0.592291
## [45] val-mlogloss:0.591669
## [46] val-mlogloss:0.591511
## [47] val-mlogloss:0.591226
## [48] val-mlogloss:0.590998
## [49] val-mlogloss:0.590114
## [50] val-mlogloss:0.589879
## [51] val-mlogloss:0.589760
## [52] val-mlogloss:0.589245
## [53] val-mlogloss:0.589075
## [54] val-mlogloss:0.588898
## [55] val-mlogloss:0.588851
## [56] val-mlogloss:0.588768
## [57] val-mlogloss:0.588583
## [58] val-mlogloss:0.588290
## [59] val-mlogloss:0.588096
## [60] val-mlogloss:0.588068
## [61] val-mlogloss:0.587981
## [62] val-mlogloss:0.587833
## [63] val-mlogloss:0.587991
## [64] val-mlogloss:0.587849
## [65] val-mlogloss:0.588040
## Stopping. Best iteration: 63

importance <- xgb.importance(feature_names = colnames(d_train), model = xgb_model)
xgb.plot.importance(importance)

#plot(varImp(xgb_model))

How well does the model predict?

val_prob <- matrix(predict(xgb_model, xgb_d_val), ncol = 3, byrow=T) %>% data.frame()
colnames(val_prob) <- c("low", "medium", "high")
val_prob <- data.frame(high = val_prob$high, low=val_prob$low, medium=val_prob$medium)
val_label <- probs_to_label(val_prob)$label
confusionMatrix(d_val$interest_level, val_label)

## Confusion Matrix and Statistics
## 
##           Reference
## Prediction high  low medium
##     high    224  267    468
##     low      43 7926    602
##     medium  149 1663    995
## 
## Overall Statistics
##                                          
##                Accuracy : 0.7413         
##                  95% CI : (0.7334, 0.749)
##     No Information Rate : 0.7989         
##     P-Value [Acc > NIR] : 1              
##                                          
##                   Kappa : 0.36           
##  Mcnemar's Test P-Value : <2e-16         
## 
## Statistics by Class:
## 
##                      Class: high Class: low Class: medium
## Sensitivity              0.53846     0.8042       0.48184
## Specificity              0.93834     0.7400       0.82360
## Pos Pred Value           0.23358     0.9247       0.35447
## Neg Pred Value           0.98313     0.4875       0.88772
## Prevalence               0.03372     0.7989       0.16738
## Detection Rate           0.01816     0.6425       0.08065
## Detection Prevalence     0.07773     0.6947       0.22753
## Balanced Accuracy        0.73840     0.7721       0.65272

sprintf("log loss = %.4f", logloss(d_val$interest_level, val_prob))

## [1] "log loss = 0.5880"

Model is predicting high & low better than medium, even though high has much lower prevalence (only 4%). But this makes sense as we found in our prior analysis that there are few features that distinguish high from low and medium. Might perform even better with one vs. all.